Improved (e.g., high-throughput, low-noise, and/or low-artifact) X-ray Microscopy images are achieved using a deep neural network trained via an accessible workflow. The workflow involves selection of a desired improvement factor (x), which is used to automatically partition supplied data into two or more subsets for neural network training. The neural network is trained by generating reconstructed volumes for each of the subsets. The neural network can be trained to take projection images or reconstructed volumes as input and output improved projection images or improved reconstructed volumes as output, respectively. Once trained, the neural network can be applied to the training data and/or subsequent data—optionally collected at a higher throughput—to ultimately achieve improved de-noising and/or other artifact reduction in the reconstructed volume.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
2. The method of claim 1, wherein the additional imaging data is of the sample acquired using the additional parameters, and wherein the additional parameters are selected to achieve a greater throughput than the training parameters.
This invention relates to imaging systems and methods for analyzing samples, particularly in applications requiring high-throughput imaging. The problem addressed is the trade-off between imaging quality and speed, where conventional systems often prioritize one at the expense of the other. The invention improves upon prior art by dynamically adjusting imaging parameters to balance throughput and accuracy. The method involves acquiring initial imaging data of a sample using a set of training parameters optimized for high-quality imaging. This data is then used to train a machine learning model to predict the sample's properties. After training, additional imaging data of the same sample is acquired using different parameters specifically selected to maximize throughput, rather than quality. The machine learning model processes this additional data to refine its predictions, allowing for faster analysis without sacrificing accuracy. The system may also include a feedback loop where the model iteratively adjusts the imaging parameters based on the results, further optimizing efficiency. The key innovation is the use of a machine learning model to compensate for lower-quality, high-throughput imaging data, enabling faster sample analysis while maintaining reliable results. This approach is particularly useful in fields like microscopy, medical imaging, and industrial inspection, where speed and accuracy are both critical.
3. The method of claim 1, wherein the training parameters are associated with a training number of projections, wherein the additional parameters are associated with an additional number of projections that is smaller than the training number of projections; and wherein the method further comprises determining the additional number of projections using the training number of projections and the improvement selection.
This invention relates to a method for optimizing the number of projections in a training process, particularly in computational imaging or machine learning applications where projection data is used to reconstruct or analyze images. The problem addressed is the computational inefficiency of using a large number of projections during training, which can be unnecessary for achieving desired performance in subsequent applications. The method involves training a system using a set of training parameters associated with a training number of projections. These projections are typically used to generate or process image data during the training phase. The system is then adapted for use with a reduced number of projections, where the additional parameters for this reduced-projection system are derived from the training parameters. The additional number of projections is smaller than the training number, making the system more efficient in terms of computational resources and data acquisition time. A key aspect of the method is determining the additional number of projections based on the training number and an improvement selection. The improvement selection likely refers to a criterion or metric that balances the trade-off between projection reduction and performance degradation. This allows the system to be optimized for specific applications where fewer projections are sufficient, such as in real-time imaging or low-resource environments. The method ensures that the reduced-projection system retains acceptable accuracy or quality while minimizing the number of projections required.
4. The method of claim 3, wherein the improvement selection is an improvement factor, and wherein determining the additional number of projections includes dividing the training number of projections by the improvement factor.
This invention relates to a method for optimizing the number of projections in a training process, particularly in machine learning or signal processing applications where projection-based techniques are used. The problem addressed is the computational inefficiency of determining an optimal number of projections during training, which can lead to excessive processing time or suboptimal performance. The method involves selecting an improvement factor, which quantifies the desired reduction in the number of projections needed to achieve a target performance level. The improvement factor is used to determine an additional number of projections by dividing a predefined training number of projections by this factor. This calculation adjusts the projection count dynamically, allowing the system to balance computational efficiency with accuracy. The method builds on a prior step of determining a training number of projections, which is the initial count used during the training phase. The improvement factor is applied to this training number to derive the additional projections, ensuring that the total number of projections remains within an efficient range while maintaining performance. This approach is particularly useful in high-dimensional data processing, where reducing the number of projections can significantly decrease computational overhead without sacrificing accuracy. The method is adaptable to various projection-based algorithms, including those used in dimensionality reduction, feature extraction, or signal reconstruction.
5. The method of claim 1, wherein partitioning the training data into the plurality of unique training subsets using the improvement selection includes partitioning the training data into a number of unique training subsets, wherein the number of unique training subsets is selected using the improvement selection.
This invention relates to machine learning model training, specifically improving the efficiency and effectiveness of partitioning training data into subsets for model optimization. The problem addressed is the challenge of selecting an optimal number of training subsets to balance computational efficiency and model performance. Traditional methods often rely on fixed or arbitrary partitioning schemes, which may not adapt to the specific characteristics of the training data or the model's learning dynamics. The invention describes a method for partitioning training data into multiple unique subsets, where the number of subsets is dynamically determined based on an improvement selection process. This process evaluates the trade-offs between the number of subsets and the resulting model improvements, such as accuracy or convergence speed. By adaptively selecting the subset count, the method ensures that the partitioning aligns with the model's learning requirements, avoiding unnecessary computational overhead or suboptimal performance. The partitioning can be applied in various machine learning contexts, including but not limited to gradient-based optimization, ensemble learning, or distributed training frameworks. The improvement selection may involve metrics like validation performance, training time, or resource utilization, allowing the method to optimize for different objectives. This adaptive approach enhances the flexibility and scalability of model training processes.
6. The method of claim 5, wherein the training data includes imaging data for a plurality of acquisitions, and wherein partitioning the training data into the number of unique training subsets using the improvement selection includes associating imaging data for sequential acquisitions of the plurality of acquisitions to alternate training subsets of the plurality of unique training subsets.
This invention relates to a method for training machine learning models using partitioned imaging data from multiple acquisitions. The problem addressed is improving model training efficiency and robustness by systematically distributing sequential imaging data across multiple training subsets. The method involves partitioning training data into unique subsets based on an improvement selection criterion. Specifically, imaging data from sequential acquisitions is alternately assigned to different training subsets. This ensures that data from consecutive acquisitions is distributed across multiple subsets, reducing bias and improving generalization. The approach leverages the temporal or sequential nature of the imaging data to enhance model performance. The partitioning process may involve iterative refinement to optimize the distribution of data across subsets. This method is particularly useful in medical imaging, remote sensing, or other fields where sequential data acquisition is common, ensuring that models trained on such data are robust and less prone to overfitting. The technique improves the reliability of machine learning models by ensuring diverse and balanced training subsets.
7. The method of claim 6, wherein sequential acquisitions of the plurality of acquisitions are acquired at different angles with respect to the sample.
This invention relates to a method for acquiring multiple images or measurements of a sample from different angles to improve imaging or analysis. The method involves capturing a sequence of acquisitions, where each acquisition is taken at a distinct angle relative to the sample. By varying the angle between successive acquisitions, the method enables enhanced visualization or characterization of the sample's structure, composition, or other properties. This approach is particularly useful in fields such as microscopy, tomography, or non-destructive testing, where capturing data from multiple perspectives improves accuracy and resolution. The method may involve rotating the sample or the imaging device between acquisitions to achieve the desired angular variation. The resulting set of angled acquisitions can be processed together to reconstruct a three-dimensional representation of the sample or to extract detailed information that would not be visible from a single angle. The technique helps overcome limitations in conventional single-angle imaging, such as shadowing, occlusion, or insufficient depth information, by providing a more comprehensive dataset for analysis.
8. The method of claim 7, wherein each sequential acquisition is angularly offset from a previous acquisition by an angle determined using golden ratio angle determination techniques.
This invention relates to a method for acquiring data in a sequential manner, particularly in imaging or sensing applications where angular positioning is critical. The method addresses the challenge of optimizing data acquisition by ensuring that each subsequent acquisition is angularly offset from the previous one using a specific mathematical approach. The golden ratio angle determination technique is employed to calculate the angular offset between sequential acquisitions. This technique leverages the golden ratio, a mathematical constant approximately equal to 1.618, to distribute acquisition angles in a way that maximizes coverage and minimizes redundancy. The method ensures that each new acquisition angle is positioned at an optimal distance from prior angles, enhancing the efficiency and accuracy of the data collection process. This approach is particularly useful in applications such as medical imaging, satellite imaging, or any system requiring precise angular sampling. By using the golden ratio, the method avoids clustering of data points and ensures uniform distribution, leading to improved data quality and reduced computational overhead. The technique can be applied in various fields where angular positioning plays a crucial role in data acquisition.
9. The method of claim 6, wherein groups of sequential acquisitions of the plurality of acquisitions are acquired at common angles, and wherein each sequential acquisition of a group of sequential acquisitions is associated with a respective one of the plurality of unique training subsets.
This invention relates to a method for acquiring and processing data from multiple acquisitions, particularly in imaging or sensing applications where data is collected at different angles. The problem addressed is improving the accuracy and efficiency of data processing by leveraging structured acquisition patterns and training subsets. The method involves acquiring a plurality of data acquisitions, where each acquisition is taken at a specific angle. These acquisitions are grouped into sets of sequential acquisitions, with each group acquired at a common angle. Within each group, individual acquisitions are associated with unique training subsets. These training subsets are used to refine or validate the data processing, ensuring that each acquisition benefits from tailored training data. The structured grouping and association with training subsets enhance the consistency and reliability of the processed data, particularly in applications like medical imaging, radar sensing, or other fields where angular data acquisition is critical. The method optimizes data collection and processing by systematically linking acquisitions to their respective training subsets, improving overall system performance.
10. The method of claim 1, further comprising receiving a region of interest (ROI) selection, wherein partitioning the training data into the plurality of unique training subsets includes using the ROI selection.
The invention relates to machine learning systems that partition training data into subsets for model training, particularly in applications where specific regions of interest (ROIs) are critical. The problem addressed is the need to improve model performance by focusing training on relevant data regions, rather than processing all available data uniformly. Traditional methods may waste computational resources on irrelevant data or fail to adequately emphasize important regions. The method involves partitioning training data into multiple unique subsets based on a user-defined ROI selection. The ROI selection identifies specific areas within the data that are prioritized for training. By incorporating this selection, the partitioning process ensures that the subsets are tailored to emphasize the ROI, improving model accuracy and efficiency. The method may also include preprocessing steps to prepare the data for partitioning, such as normalization or feature extraction, and may apply different partitioning strategies (e.g., random sampling, stratified sampling) to create diverse training subsets. The resulting subsets are then used to train a machine learning model, with the ROI-informed partitioning enhancing the model's ability to learn from the most relevant data. This approach is particularly useful in fields like medical imaging, where certain anatomical regions require more detailed analysis.
15. The method of claim 13, wherein determining that the trained neural network is insufficient is performed automatically using a machine learning classifier.
A system and method for evaluating the performance of a trained neural network involves automatically assessing whether the network is insufficient using a machine learning classifier. The neural network is initially trained on a dataset to perform a specific task, such as classification, regression, or pattern recognition. After training, the system evaluates the network's performance using predefined metrics, such as accuracy, precision, recall, or loss. If the performance falls below a threshold, the system determines that the neural network is insufficient. This determination is made automatically by a machine learning classifier, which analyzes the performance metrics and other relevant data to decide whether the network meets acceptable standards. The classifier may be trained on historical data to recognize patterns indicative of insufficient performance. If the network is deemed insufficient, the system may trigger retraining, adjustment of hyperparameters, or other corrective actions to improve performance. This approach ensures that neural networks are continuously monitored and optimized for reliability and accuracy in their intended applications.
16. The method of claim 1, further comprising applying an angle-dependent weighting mask to the training data.
This invention relates to machine learning systems, specifically improving training data processing for enhanced model performance. The core problem addressed is the variability in training data quality, where certain data points may be more informative or relevant than others, leading to suboptimal model training. The solution involves applying an angle-dependent weighting mask to the training data, which adjusts the influence of individual data points based on their angular relationships within the feature space. This weighting mechanism prioritizes data points that contribute more significantly to the model's learning process, thereby improving accuracy and robustness. The method integrates with a broader training framework that includes data preprocessing, feature extraction, and model optimization. The angle-dependent weighting mask dynamically adjusts weights during training, ensuring that the model focuses on the most relevant data points while mitigating the impact of noisy or less informative samples. This approach enhances the model's ability to generalize from the training data, leading to better performance on unseen data. The invention is particularly useful in applications where data quality varies significantly, such as image recognition, natural language processing, and other domains requiring high-precision machine learning models.
17. The method of claim 1, further comprising truncating each of the CT training reconstructed volumes in a Z direction prior to training the neural network.
This invention relates to medical imaging, specifically improving the training of neural networks for computed tomography (CT) image reconstruction. The problem addressed is the computational inefficiency and potential inaccuracies in training neural networks on full CT volumes, which can be large and contain irrelevant or redundant data. The method involves preprocessing CT training data by truncating each reconstructed CT volume in the Z direction (along the longitudinal axis) before training the neural network. This truncation reduces the size of the training data, making the training process faster and more efficient. The neural network is then trained on these truncated volumes to improve its ability to reconstruct high-quality CT images from raw projection data. The truncation step ensures that only the most relevant portions of the CT volumes are used for training, which can enhance the neural network's performance by focusing on the most informative regions. This approach is particularly useful in medical imaging applications where fast and accurate image reconstruction is critical for diagnosis and treatment planning. The method can be applied to various neural network architectures designed for CT image reconstruction, improving their efficiency and effectiveness.
21. A system for processing X-ray imaging data, the system including a control system configured to implement the method of claim 1.
The system processes X-ray imaging data to enhance image quality and diagnostic accuracy. X-ray imaging often suffers from noise, artifacts, and low contrast, which can obscure critical anatomical details and reduce diagnostic reliability. The system addresses these issues by implementing an advanced processing method that includes noise reduction, artifact correction, and contrast enhancement. The control system receives raw X-ray imaging data from an imaging device and applies a series of algorithms to filter out noise, remove artifacts caused by patient movement or equipment limitations, and improve contrast to highlight relevant structures. The processed images are then output for display or further analysis. The system may also include calibration and feedback mechanisms to adapt processing parameters based on imaging conditions or user preferences. By automating these enhancements, the system reduces the need for manual adjustments and ensures consistent, high-quality imaging results across different clinical settings. The technology is particularly useful in medical diagnostics, where accurate and clear X-ray images are essential for effective patient care.
22. A computer program product embodied on a non-transitory computer readable medium and comprising instructions which, when executed by a computer, cause the computer to carry out the method of claim 1.
A system and method for automated data processing involves analyzing input data to identify patterns or anomalies using machine learning techniques. The system receives input data from one or more sources, preprocesses the data to remove noise and normalize values, and applies machine learning algorithms to detect patterns, trends, or anomalies. The system then generates output data based on the analysis, which may include alerts, reports, or recommendations. The machine learning models are trained using historical data to improve accuracy over time. The system may also include a user interface for configuring parameters, reviewing results, and adjusting the analysis process. The method ensures efficient and accurate data processing by leveraging automated techniques to reduce manual intervention and improve decision-making. The system can be applied in various domains, such as fraud detection, predictive maintenance, or quality control, where identifying patterns or anomalies in data is critical. The computer program product stores the instructions for executing this method on a non-transitory computer-readable medium, enabling deployment across different computing environments.
25. The method of claim 23, wherein the additional imaging data is of the sample acquired using the additional parameters, and wherein the additional parameters are selected to achieve a greater throughput than the training parameters.
This invention relates to imaging systems and methods for analyzing samples, particularly in applications requiring high throughput. The problem addressed is the trade-off between imaging quality and speed, where conventional systems must balance resolution, accuracy, and processing time. The invention improves upon prior methods by dynamically adjusting imaging parameters to optimize throughput without sacrificing critical analytical performance. The method involves acquiring initial imaging data of a sample using a set of training parameters, which are optimized for accuracy but may limit speed. Additional imaging data is then collected using different parameters specifically chosen to increase throughput. These additional parameters may include faster scanning rates, lower resolution settings, or reduced data acquisition intervals, depending on the sample and analysis requirements. The system processes both datasets to extract relevant information, leveraging the high-quality training data for calibration or reference while using the faster-acquired data for rapid analysis. This approach allows the system to maintain analytical rigor where needed while significantly increasing overall throughput for large-scale or time-sensitive applications. The method is particularly useful in fields like medical diagnostics, industrial quality control, or environmental monitoring, where both speed and accuracy are critical. The invention ensures that the additional imaging data, though acquired under different conditions, remains sufficiently reliable for its intended purpose.
26. The method of claim 23, wherein the training parameters are associated with a training number of projections, wherein the additional parameters are associated with an additional number of projections that is smaller than the training number of projections; and wherein the method further comprises determining the additional number of projections using the training number of projections and the improvement selection.
This invention relates to a method for optimizing the number of projections in a machine learning or signal processing system, particularly for tasks like image reconstruction or data compression. The problem addressed is the computational inefficiency of using a fixed, often excessive, number of projections during both training and deployment phases, which can lead to unnecessary resource consumption without significant performance gains. The method involves a two-phase approach: a training phase and a deployment phase. During training, a system is trained using a training number of projections, which is typically a larger set designed to capture detailed information. The training process generates training parameters that define the learned model. In the deployment phase, the system operates with an additional number of projections that is smaller than the training number, reducing computational overhead. The additional number of projections is determined based on the training number and an improvement selection, which evaluates the trade-off between performance and efficiency. This selection may involve metrics like reconstruction accuracy, computational cost, or other task-specific criteria. By dynamically adjusting the number of projections, the method balances performance and resource usage, making it suitable for real-time or resource-constrained applications.
27. The method of claim 26, wherein the improvement selection is an improvement factor, and wherein determining the additional number of projections includes dividing the training number of projections by the improvement factor.
This invention relates to a method for optimizing the number of projections in a training process, particularly in machine learning or signal processing applications where computational efficiency is critical. The problem addressed is the need to reduce the number of projections required during training while maintaining or improving performance, thereby saving computational resources and time. The method involves selecting an improvement factor, which quantifies the desired reduction in the number of projections. The training process initially uses a predefined number of projections, and the improvement factor is applied to determine the additional number of projections needed. Specifically, the additional number of projections is calculated by dividing the original training number of projections by the improvement factor. This adjustment allows the system to dynamically adapt the number of projections based on performance metrics or other criteria, ensuring efficient resource utilization without compromising accuracy. The method may also include steps for evaluating the effectiveness of the selected improvement factor, such as comparing performance metrics before and after the adjustment. If the improvement factor leads to suboptimal results, it can be refined or replaced with a different value. This iterative approach ensures that the system continuously optimizes the number of projections for the best balance between computational efficiency and performance. The invention is particularly useful in applications where real-time processing or large-scale data analysis is required, such as in image recognition, natural language processing, or sensor data analysis.
28. The method of claim 23, wherein partitioning the training data into the plurality of unique training subsets using the improvement selection includes partitioning the training data into a number of unique training subsets, wherein the number of unique training subsets is selected using the improvement selection.
This invention relates to machine learning model training, specifically improving the efficiency and effectiveness of partitioning training data into subsets for model optimization. The problem addressed is the challenge of selecting an optimal number of training subsets to balance computational efficiency and model performance during iterative training processes. Traditional approaches often rely on fixed or arbitrary subset sizes, which may not adapt to the specific characteristics of the training data or the model's learning dynamics. The invention describes a method for partitioning training data into multiple unique subsets, where the number of subsets is dynamically determined based on an improvement selection criterion. This criterion evaluates the trade-offs between the granularity of subset partitioning and the computational resources required for training. By adjusting the number of subsets in response to observed improvements in model performance, the method ensures that training resources are allocated optimally. The partitioning process may involve techniques such as stratified sampling, clustering, or other data-driven approaches to create subsets that are representative of the overall training data. The improvement selection criterion can be based on metrics like validation accuracy, loss function convergence, or computational cost, allowing the method to adapt to different training scenarios. This dynamic partitioning approach enhances the efficiency of model training by reducing unnecessary computations while maintaining or improving model accuracy.
29. The method of claim 28, wherein the training data includes imaging data for a plurality of acquisitions, and wherein partitioning the training data into the number of unique training subsets using the improvement selection includes associating imaging data for sequential acquisitions of the plurality of acquisitions to alternate training subsets of the plurality of unique training subsets.
This invention relates to machine learning training methods for medical imaging, specifically addressing the challenge of improving model performance by optimizing the partitioning of training data. The method involves training a machine learning model using imaging data from multiple acquisitions, where the data is divided into distinct training subsets. The key innovation lies in the partitioning strategy, which ensures that imaging data from sequential acquisitions is distributed across different training subsets. This approach prevents overfitting and enhances generalization by reducing dependencies between consecutive acquisitions. The method further includes an improvement selection process to evaluate and refine the partitioning, ensuring optimal data distribution. By alternating sequential acquisition data across subsets, the model learns from diverse inputs, improving robustness and accuracy in medical imaging applications. The technique is particularly useful in scenarios where sequential imaging data may contain correlated artifacts or biases, such as in time-series medical scans or longitudinal studies. The method dynamically adjusts the partitioning based on performance metrics, ensuring continuous refinement of the training process. This approach enhances the reliability of machine learning models in medical diagnostics and imaging analysis.
30. The method of claim 29, wherein sequential acquisitions of the plurality of acquisitions are acquired at different angles with respect to the sample.
This invention relates to a method for acquiring multiple images or measurements of a sample from different angles to improve imaging or analysis. The method involves capturing a sequence of acquisitions, where each acquisition is taken at a distinct angle relative to the sample. By varying the angle between successive acquisitions, the method enables enhanced visualization or characterization of the sample, which can be particularly useful in applications such as microscopy, tomography, or non-destructive testing. The different angles allow for capturing different perspectives or cross-sectional views, which can be combined to reconstruct a more comprehensive representation of the sample's structure or properties. This approach helps overcome limitations in single-angle imaging, such as shadowing, occlusion, or insufficient depth information, by providing a multi-angle dataset that can be processed to extract additional details or improve accuracy. The method may be applied in various imaging modalities, including optical, electron, or X-ray imaging, depending on the specific requirements of the analysis. The sequential acquisition at different angles ensures that the sample is systematically scanned from multiple directions, enabling advanced techniques like 3D reconstruction, tomographic imaging, or angular-dependent analysis.
31. The method of claim 30 wherein each sequential acquisition is angularly offset from a previous acquisition by an angle determined using golden ratio angle determination techniques.
This invention relates to a method for acquiring data in a sequential manner, where each subsequent acquisition is angularly offset from the previous one by an angle derived using golden ratio angle determination techniques. The method is designed to optimize data collection by ensuring uniform angular distribution of acquisition points, which is particularly useful in applications requiring precise spatial sampling, such as imaging, sensing, or measurement systems. The golden ratio, approximately 137.5 degrees, is used to determine the angular offset between sequential acquisitions, ensuring that each new acquisition is positioned in a way that minimizes redundancy and maximizes coverage of the sampling space. This approach improves the efficiency and accuracy of data acquisition by distributing sampling points more evenly compared to traditional linear or arbitrary spacing methods. The technique is applicable in various fields, including medical imaging, remote sensing, and industrial inspection, where uniform sampling is critical for high-quality data reconstruction. By leveraging the mathematical properties of the golden ratio, the method ensures that each acquisition contributes uniquely to the overall dataset, reducing gaps and overlaps in the collected data. This results in more reliable and comprehensive data representation, enhancing the performance of subsequent analysis and interpretation processes.
32. The method of claim 30, wherein groups of sequential acquisitions of the plurality of acquisitions are acquired at common angles, and wherein each sequential acquisition of a group of sequential acquisitions is associated with a respective one of the plurality of unique training subsets.
This invention relates to a method for acquiring and processing data from multiple sequential acquisitions, particularly in imaging or sensing applications where data is collected at different angles or positions. The method addresses the challenge of efficiently training and validating models or algorithms by dividing the acquired data into unique subsets for training and testing purposes. The method involves acquiring a plurality of data acquisitions, where each acquisition is taken at a specific angle or position. These acquisitions are grouped such that each group consists of sequential acquisitions taken at the same angle. Within each group, individual acquisitions are associated with different unique training subsets. This ensures that data from the same angle is distributed across multiple training subsets, preventing overfitting and improving the robustness of the trained model. By structuring the acquisitions in this way, the method enables comprehensive training and validation of models using diverse data while maintaining consistency within each training subset. This approach is particularly useful in applications such as medical imaging, remote sensing, or industrial inspection, where data is collected from multiple angles or positions and must be analyzed for accuracy and reliability. The method ensures that the training process leverages the full range of acquired data while avoiding biases that could arise from using the same angle data in both training and testing.
33. The method of claim 23, further comprising receiving a region of interest (ROI) selection, wherein partitioning the training data into the plurality of unique training subsets includes using the ROI selection.
This invention relates to machine learning systems that partition training data for model training. The problem addressed is the need to efficiently and accurately segment training datasets to improve model performance, particularly when certain regions of interest (ROIs) are prioritized. The method involves partitioning training data into multiple unique subsets, where each subset is used to train a separate model or a portion of a model. The partitioning process incorporates a user-defined ROI selection, ensuring that the subsets are tailored to focus on specific areas of the data that are critical for the application. This approach enhances model accuracy and efficiency by directing computational resources toward the most relevant data regions. The method may also include preprocessing steps to prepare the training data, such as normalization or feature extraction, before partitioning. The use of ROI selection allows for adaptive training, where the model can be fine-tuned based on the importance of different data regions. This technique is particularly useful in applications like medical imaging, autonomous driving, or any domain where certain data regions are more significant than others. The invention improves upon prior methods by dynamically incorporating user-defined priorities into the training process, leading to more effective and specialized machine learning models.
38. The method of claim 36, wherein determining that the trained neural network is insufficient is performed automatically using a machine learning classifier.
A system and method for evaluating the performance of a trained neural network involves automatically determining whether the network is insufficient using a machine learning classifier. The neural network is initially trained on a dataset to perform a specific task, such as classification, regression, or pattern recognition. After training, the system assesses the network's performance by applying it to a validation dataset and measuring its accuracy, precision, recall, or other relevant metrics. If the performance falls below a predefined threshold, the system automatically identifies the network as insufficient. This determination is made using a machine learning classifier that analyzes the network's performance metrics, architecture, or other relevant features to predict whether the network meets acceptable standards. The classifier may be trained on historical data from previous neural network evaluations to improve accuracy. If the network is deemed insufficient, the system may trigger retraining, hyperparameter tuning, or other corrective actions to improve performance. This approach automates the evaluation process, reducing manual intervention and ensuring consistent assessment of neural network quality. The method is applicable in various domains, including computer vision, natural language processing, and predictive analytics, where reliable neural network performance is critical.
39. The method of claim 23, further comprising applying an angle-dependent weighting mask to the training data.
This invention relates to a method for improving the accuracy of a machine learning model by applying an angle-dependent weighting mask to training data. The method addresses the challenge of training models on datasets where certain angles or orientations of data points are underrepresented or overrepresented, leading to biased or inaccurate predictions. The angle-dependent weighting mask adjusts the influence of training samples based on their angular orientation, ensuring that the model learns from a more balanced distribution of angles. This technique is particularly useful in applications such as computer vision, where objects may appear at various angles, and in robotics, where orientation-dependent tasks require precise modeling. The weighting mask is applied during the training phase, dynamically adjusting the contribution of each training sample to the model's learning process. This ensures that the model generalizes better across different angles, improving its robustness and performance in real-world scenarios. The method may be combined with other data augmentation techniques to further enhance model accuracy. By incorporating angle-dependent weighting, the invention provides a solution for training models that perform consistently across a wide range of orientations.
40. The method of claim 23, wherein training the neural network further includes truncating each of the CT training reconstructed volumes in a Z direction prior to evaluating the neural network.
This invention relates to medical imaging, specifically improving neural network training for reconstructing computed tomography (CT) volumes. The problem addressed is the computational inefficiency and potential inaccuracies in training neural networks for CT image reconstruction, particularly when processing large three-dimensional datasets. The solution involves a preprocessing step where each CT training reconstructed volume is truncated in the Z direction (typically the axial or slice direction) before being evaluated by the neural network. This truncation reduces the data size, making the training process more efficient while maintaining reconstruction quality. The neural network is trained using these truncated volumes to learn patterns and features relevant to CT image reconstruction. The method may also include other preprocessing steps, such as normalizing the data or applying noise reduction techniques, to further enhance training performance. The truncated volumes are used to optimize the neural network's parameters, ensuring accurate and efficient reconstruction of CT images during inference. This approach is particularly useful in clinical settings where fast and reliable CT image reconstruction is critical.
44. A system for processing X-ray imaging data, the system including a control system configured to implement the method of claim 23.
The system processes X-ray imaging data to enhance image quality and diagnostic accuracy. X-ray imaging often suffers from noise, artifacts, and low contrast, which can obscure critical anatomical details and reduce diagnostic reliability. The system addresses these challenges by implementing an advanced processing method that combines noise reduction, artifact correction, and contrast enhancement techniques. The control system receives raw X-ray imaging data from an imaging device and applies a series of computational algorithms to improve the signal-to-noise ratio, remove artifacts caused by patient movement or equipment limitations, and enhance contrast to highlight relevant structures. The processing method includes adaptive filtering to preserve fine details while reducing noise, iterative reconstruction to correct distortions, and dynamic contrast adjustment to optimize visibility of different tissue types. The system may also incorporate machine learning models trained on clinical datasets to further refine image quality based on specific diagnostic requirements. The output is a high-quality X-ray image suitable for accurate diagnosis, reducing the need for retakes and improving patient outcomes. The system is designed for integration with existing radiology workflows, ensuring seamless adoption in clinical settings.
45. A computer program product embodied on a non-transitory computer readable medium and comprising instructions which, when executed by a computer, cause the computer to carry out the method of claim 23.
A system and method for optimizing data processing in a distributed computing environment addresses inefficiencies in task allocation and resource utilization. The invention provides a dynamic scheduling mechanism that analyzes workload characteristics, system resources, and network conditions to assign tasks to processing nodes in a way that minimizes latency and maximizes throughput. The system monitors performance metrics in real-time, such as task completion times, resource availability, and network bandwidth, to adjust task distribution dynamically. It also includes a predictive model that forecasts future workload demands based on historical data and current trends, allowing proactive resource allocation. The method further incorporates fault tolerance by detecting and rerouting tasks from failed or underperforming nodes to ensure uninterrupted processing. Additionally, the system supports heterogeneous computing environments, enabling seamless integration of different hardware architectures and software frameworks. The invention improves overall system efficiency by reducing idle time, balancing load across nodes, and optimizing data transfer between components. This approach is particularly useful in large-scale distributed systems, such as cloud computing platforms, where efficient resource management is critical for performance and cost-effectiveness.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
July 9, 2021
April 30, 2024
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.